Robbyant · 2026-07-09 · notable
LingBot-Video — Apache-2.0 30B-A3B MoE video model for embodied AI
LingBot-Video is an open-source Mixture-of-Experts video generation model trained on more than 70,000 hours of embodied data, released July 9 by Robbyant. Weights ship as a 1.3B dense variant and a 30B-total 3B-active MoE with Apache-2.0 code.
Robbyant open-sources a 30B Mixture-of-Experts video model aimed at giving robots a shared, physically plausible world imagination.
Quick facts
| Maker | Robbyant |
|---|---|
| Variants | LingBot-Video-Dense 1.3B and LingBot-Video-MoE 30B-A3B |
| License | Apache-2.0 |
| Training data | 70,000+ hours of embodied video |
| Paper | arXiv 2607.07675 |
| Assets | Code, weights, and rewriting tools |
What is it?
LingBot-Video is a Mixture-of-Experts video generation model Robbyant released on July 9 with Apache-2.0 code, weights, and a technical report on arXiv (2607.07675). LingBot-Video ships as a 1.3B parameter dense variant for text-to-image, text-to-video, and text-image-to-video, plus a 30B-total 3B-active MoE with a refiner for stronger generation and rewriting.
How does it work?
The model scales Mixture-of-Experts from scratch, so only 3B of the 30B parameters are active per token, which Robbyant reports gives roughly 3× faster inference than a dense model at the same quality. Training mixes web video with 70,000+ hours of embodied data and uses a multi-reward system that scores outputs for aesthetics, physical rationality, and task completion together, rather than aesthetics alone.
Why does it matter?
Video-first world models are becoming the shared imagination layer robotics teams pair with a policy, and most strong ones are closed. LingBot-Video is the first open-source large-scale MoE video model built explicitly for embodied intelligence, and the release drew 495 upvotes on Hugging Face Papers on launch day. Apache-2.0 weights mean labs and startups can fine-tune the model without a licence conversation.
Who is it for?
Robotics and embodied-AI researchers, plus video-generation teams looking at MoE architectures with open weights.
Frequently asked questions
- What is LingBot-Video and how is it different from a normal video model?
- LingBot-Video is a video generation model Robbyant open-sourced on July 9 that is trained for embodied intelligence — the video has to look right and be physically plausible so it can drive robot policies. LingBot-Video mixes web video with more than 70,000 hours of embodied data and rewards outputs for physical rationality and task completion, not only aesthetics.
- Which sizes of LingBot-Video are released?
- Robbyant ships two variants of LingBot-Video: a 1.3B parameter dense model for text-to-image, text-to-video, and text-image-to-video, and a Mixture-of-Experts version with about 30 billion total parameters and 3 billion active per token, plus a refiner. Both are Apache-2.0 with weights on Hugging Face and code on GitHub.
- Can I use LingBot-Video commercially?
- LingBot-Video is released under Apache-2.0, so commercial use is allowed as long as the license and notices are preserved. Robbyant publishes the code, weights, and rewriting tools together, and paper 2607.07675 documents the training recipe. There are no non-commercial or research-only carve-outs on the model card as released today.
Try it
Clone github.com/Robbyant/lingbot-video and pull the 30B-A3B checkpoint from huggingface.co/robbyant/lingbot-video-moe-30b-a3b.